Embodied AI refers to intelligence that possesses a physical body, perceives the environment, and interacts with it through action—it does not merely "learn in data" but also "acts in the world." Unlike pure software intelligence, embodied agents must handle real physical constraints: gravity, friction, latency, sensor noise, and the consequences of action failures.

The conceptual origins of Embodied AI lie in the behavior-based robotics of the 1980s by Brooks, which emphasized "perception-action" coupling rather than first building world models; the term "embodied AI" became fixed around 2018–2019 with the rise of deep reinforcement learning and simulation platforms.

A representative infrastructure for Embodied AI is Habitat. In 2019, institutions such as Meta AI released the Habitat simulation platform in the paper Habitat: A Platform for Embodied AI Research, turning "training embodied agents in virtual environments" into a reproducible experimental paradigm.

The research framework for Embodied AI consists of four components: embodied perception, embodied interaction, embodied agents, and sim-to-real transfer—this division comes from the 2024 survey by Liu et al., Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI. The survey also notes that multimodal large models and world models are becoming powerful architectures for embodied agents, with simulation platforms like Habitat, Isaac Sim, and MuJoCo providing low-cost training grounds; data is another bottleneck, as real robot data is expensive, with teleoperation collection and simulation synthesis serving as the primary sources. Embodied AI serves as a bridge connecting digital intelligence to the physical world (robotics, smart manufacturing, medical rehabilitation) and is viewed as a key step toward general intelligence; the biggest hurdle remains the sim-to-real gap—skills learned in simulation often require re-tuning in the real world.

[JIALIU_0]